Tae posted:

How expensive are these mansions, jeez

golden bubble posted:

Towa and Roberu manage a near perfect EN/ID and JP split in viewship. Nene and Botan have some special fans. Nene is the only JP Hololive with a noticeable Spanish following. Botan is the only one with a significant Russian following.

astr0man posted:

Is this just based on what % of their chat messages are in EN vs JP, or was it actually looking at whatever info's available on people's yt/google account profiles?

https://www.reddit.com/r/Hololive/comments/l2f9en/which_member_gets_the_most_english_chat_messages/ posted:

TL;DR

No collabs, no “English only!” challenges, and no “English study/talk” streams included

No messages consisting solely of emojis, punctuation, numbers, or ‘w’ spam counted

EN / ID is anything that uses only A-Z, ES is anything that uses Latin characters but goes beyond simple A-Z (eg diacritics), RU is anything that uses Cyrillic, JP is everything else

Dataset is, in general, around the most recent ~10 streams of each member’s, with more added if needed to hit 15 hour and 50,000 message minimums (minimums not applicable to Holostars)

Graphs round to 1 decimal place and don’t show percents below 1%, so stuff doesn’t always add up to 100%

I made specific notes about Miko, Haachama, Pekora, Coco, and Towa below. Please read those first if you have a question, concern, or particular interest about any of those members’ results.

I will not be doing HoloID or HoloEN as their charts will just be a bunch of 99% or 100% EN / ID

Introduction

I’ve always been curious about the language breakdown of Holo members’ chats – who gets the most English messages, who gets the fewest, what percent of their chat is English, just how many Russian messages does Botan get, etc. – so I thought it would be a fun project to analyze the data and try to answer these questions. For this, I wrote a program that reads each of the chat messages on a stream, determines what language it is, and collates all the data, and then I graphed that data. As the images say, all in all I ended up analyzing almost 3 million chat messages, and these are the results.

Data Collection Methodology

I first had to determine exactly where to get the messages to analyze. My goal for this project was to get the language breakdown of the average stream for each member. I didn’t want the data to be skewed by content such as unique, one-off streams, especially ones that had a specific language-focus to them. To that end, I established 2 rules for determining which streams to analyze – (1) no collabs, as collabs run the risk of the other collab member’s audience too heavily influencing the chat of the streamer I was observing, and (2) no language-focused streams, in other words, no “English only!” challenges, no “English study” streams, etc. To note that rule (2) had a very minimal effect and only ended up excluding 2 Sora streams, 1 Shien stream, and 1-2 Coco streams (see below for more about Coco).

Next, I had to determine how to parse each message. The first step was a bit of preprocessing – if a message was solely numerical, an emoji, punctuation marks, only ‘w’s, or any combination of these, I discarded the message entirely and did not count it towards any individual language or towards the total number of messages, as such a message could not accurately be assigned to any individual language. Next, I had to place each message into the corresponding language bucket. In the image, I referred to the four buckets as EN / ID, JP, ES, and RU, but that isn’t 100% accurate due to the parsing algorithm I used. Here is the full definition of each bucket:

EN / ID – Any message that only uses Latin characters found in the English alphabet (A-Z). This primarily captures English and Indonesian (as both only use the 26 standard English letters), but it also can end up mistakenly capturing non-English messages from other Latin-based languages if those messages happened to not use any special letters. This may occur either because the writer was too lazy to properly write diacritics or if that particular message just happened to not contain any. The overall effect of this is that the EN / ID is very slightly over-counted, however the number of people writing unaccented Spanish, French, Italian, etc. messages in Holo members’ chats is extremely low, so the very large sample size should mostly eliminate any real bias this would cause.

ES – Any message written using Latin characters where at least 1 character is a non-English letter. This covers everything from diacritics like Spanish é and German ä to entirely new letters like Scandinavian Ø. While this bucket technically encompasses many different languages, for Holo purposes it’s mostly Spanish (and perhaps Portuguese) messages, so I have merely called the bucket “ES” for convenience.

RU – Any message written using Cyrillic characters. While there are technically many languages besides Russian that use the Cyrillic alphabet, I think it’s safe to say that the vast majority of any Cyrillic messages are going to be in Russian, so I think it’s fair to call this bucket “RU.”

JP – Any message that was not outright excluded in preprocessing and does not fall into one of the above 3 buckets. Due to the extremely large number of characters in the Japanese language, I decided to go with an exclusionary approach to determining if something was a Japanese message. This means that technically any messages not written using either Latin or Cyrillic characters get counted as JP messages. So, for example, messages in Arabic, Chinese, or Korean would end up getting counted in the JP bucket. Similar to the EN / ID bucket, due to the extremely low number of messages in those languages compared to the huge sample size of messages, the effects of this should not really be noticeable.

With all that out of the way, the last step was just deciding which individual streams to use. For this I pretty much just chose whatever the member’s most recent streams were so that I could get the most up-to-date data possible. In two specific instances, which I’ll note below, I did decide to forego a few more recent streams in favor of older streams in an attempt to get a more representative sample of that member’s average stream.

In terms of the volume of data, I used a minimum of 9 different streams per member (the exact amount varies by member based on a variety of other factors), a minimum of 15 hours of content per member, and a minimum of 50,000 chat messages for each Hololive member. Holostars had slightly laxer requirements, as they obviously get less chat messages, but I still used a minimum of 9 streams for each member.

Graphing

For the graphs, I rounded values to one decimal place. I also excluded any values below 1%, as they would be barely visible on most graphs and merely clutter up the graph. As a result, you will notice that many of the charts don’t add up to exactly 100%, due to both rounding errors and not including the small ES and RU percentages. In general, the further away from 100% the two shown numbers add to, the more ES and RU comments that member received.

Channel Specific Notes

Miko – Miko has been streaming a lot of Yakuza lately, which attracts a very Japanese-heavy chat compared to other stream content. I did include some Yakuza streams in her dataset, but I also passed over a bunch to include some earlier Minecraft streams instead in an effort to better represent the average content on her channel. It’s not as if Miko only plays Yakuza, and actually plays Minecraft quite regularly, so it didn’t really make sense to me to have, say, 8 Yakuza streams and 0 Minecraft streams in the dataset.

Haachama – I debated for a long time with myself if I should exempt Haachama from the language-specific content rule and make an active effort to include some of her English-focused streams in her dataset. She is in a very unique position among Hololive members and had previously really been making an effort to make a lot of content specifically for the English-speaking audience. However, I ultimately decided against it, as lately Haachama has not really been doing English-language content outside of collabs (her last solo English stream was on December 17), so I decided that her “average stream”, at least at the moment, is not really English-focused.

Pekora – See Miko. Same thing where I excluded some more recent Yakuza streams for earlier streams of different content. I still included some Yakuza streams, of course, and her Yakuza streams are also very long, so they’ll tend to contribute a large amount of messages. As a result, Pekora’s normal JP % (ie when she finishes streaming Yakuza so much) is probably a bit lower than the data here indicates.

Coco – This was another one that I debated a long time with myself about – should I include Coco’s meme reviews or not? On one hand, she very regularly does them on a schedule, so it can certainly be said that they’re part of her “average content.” But, on the other hand, you can argue that they are language-specific content and will create skew in the chat because of that. Ultimately, I decided to not include Coco’s meme reviews in her dataset. I can certainly see the other argument, too, though, and would not fault anyone who thinks they should be included. I had to make a decision, though, and that is what I chose.

Also on the topic of Coco, I will note that Coco’s language breakdown is very unique among Holo members. On most streams, she gets very, very few English comments – comparable to the lowest overall in Hololive. However, sometimes she randomly decides to speak mostly in English instead of Japanese on some streams, and on those streams she’ll instead get a ton of English comments, even outnumbering the Japanese ones, so this pushes her overall EN / ID percentage up to the 22.8% you see in the graph. Most members have a fairly consistent % across streams within a couple percentage points each way of their average, but Coco’s individual stream %’s instead have extremely high volatility.

Towa – There is one other bit of preprocessing to messages I did that pertains to Towa. Any message solely containing “TMT”, “TMD”, or “TCA” was excluded and not counted in any language’s bucket or in the total count of messages. This is because it’s a fairly cross-language thing to spam these letters at Towa, as you can’t type “TMT” in Japanese really (besides typing out the entire phrase which is a total hassle). Thus, counting them as EN / ID messages would be fairly misleading, as lots of the people typing them are probably actually Japanese.

If you’re curious, about 5.4% of the total messages in Towa’s chat (not counting emoji, numerical, etc. messages in the total) are some variation of “TMT.”

General Observations and Comments

Looking at the raw data, language shares tend to vary heavily with the type of content being streamed, as one might expect. Talking streams tend to get very low English shares, while gaming streams tend to get more, and singing streams the highest English share of all types of content. Among gaming streams, the exact game being streamed also seems to make a noticeable difference. Games like Fall Guys, Apex, and GTA attract many more EN / ID comments than games like Yakuza, ARK, or Mario Kart. Minecraft seems to be a fairly neutral game, with it not showing any consistent deviation one way or the other from each member’s average.

The total messages per hour chart might surprise some people, particularly the fact that Pekora is not even being in the top 5 despite getting by far the most viewers out of all HololiveJP members. (Pekora is actually 8th, if you’re curious, behind the pictured five, Ayame (#6), and Rushia (#7)). If I may offer some explanations, there are a few possible ones that I can think of. It could simply be the case that Pekora simply attracts a lot more lurkers than other members. Perhaps her stream has more mainstream popularity, where many viewers enjoy watching it for entertainment, but aren’t invested enough in the Youtube ecosystem to actually participate in chat. Another explanation might be due to a significant portion of her dataset being made up of Yakuza streams, as I noted before. Perhaps Yakuza simply does not attract many chat messages compared to other types of content, so this drags her average down. A final explanation that comes to mind centers on the way that I processed the data. I immediately discarded and didn’t count any messages which consisted solely of an emoji, and, in my experience at least, Pekora’s viewers – for whatever reason – tend to spam emojis a lot more than other channels’ viewers do. It could be the case that Pekora’s chat actually does get the overall highest messages per hour, but my algorithm simply discarded most of those messages since it was primarily focused on language parsing and counting total chat messages was just a fun side statistic.

As a final comment, I would just like to remind everyone that this is, of course, not a definitive analysis. While I tried to be as rigorous as possible in my methods, ultimately this is only an analysis using around 10 streams of each member. If you extended the dataset to 30, 50, 100, etc. streams, you may find that you suddenly come up with different numbers. Collecting just this much data took me over a week, though, so you’ll forgive me if I wasn’t about to go catalogue the last 50 streams of each member. That said, other than any specific points of interest that I noted above, I do believe that the data presented here should be fairly accurate and that any additional data collection would lead to, at most, only a few percentage points swing in either direction.

What about HoloID and HoloEN?

I considered extending my analysis to EN and ID, but ultimately after doing a couple test experiments, their chats – even for members who can speak Japanese – are almost exclusively EN / ID messages. All other buckets would likely fall below the 1% threshold, or at best be barely above it, and looking at a bunch of 99% and 100% pie charts is not very informative or interesting, so I will not be doing the same analysis for HoloID or HoloEN.

Closing

If there’s anything that I didn’t mention here that you’re curious about, whether it be about the data itself, my methodology, or whatever, feel free to ask and I’ll do my best to answer. Oh and I apologize for the (lack of) graphic design in the images. I’m good at coding / statistics, not art.

astr0man posted:

Is this just based on what % of their chat messages are in EN vs JP, or was it actually looking at whatever info's available on people's yt/google account profiles?

Taratang posted:

It's just chat messages - actual viewer metrics are not public information and rarely discussed openly (especially after the whole Taiwan incident), but Pekora did mention a little while back that she was shocked in a recent management meeting when they showed her her overseas viewership had climbed to almost 60% of her total viewers.

quote:

Also on the topic of Coco, I will note that Coco’s language breakdown is very unique among Holo members. On most streams, she gets very, very few English comments – comparable to the lowest overall in Hololive. However, sometimes she randomly decides to speak mostly in English instead of Japanese on some streams, and on those streams she’ll instead get a ton of English comments, even outnumbering the Japanese ones, so this pushes her overall EN / ID percentage up to the 22.8% you see in the graph. Most members have a fairly consistent % across streams within a couple percentage points each way of their average, but Coco’s individual stream %’s instead have extremely high volatility.

Sundae posted:

Julias posted:

That's a pretty good name/design. Kudos to your friend, and good luck to her and her foray into vtubing!

sb hermit posted:

Talorat posted:

Wait. Holy poo poo. Has the same emote fish been a shark because same this whole time.

Revolver Bunker posted:

... Sock gloves, gummy bear tarot cards, low lighting on a dark table, and her going through the list of fellow members...

Misanthropic Void posted:

If only her crimes against tarot (or cards in general) stopped there. Her shuffling and drawing methods where.... original to say the least.

Kata-Haro posted:

Nene and Polka really have some strong sibling energy going

https://www.youtube.com/watch?v=ASSmcRounNU

Jerkface posted:

I love them both, can't wait for Nene's new outfit it should be next weekend??? Soon?

Sundae posted:

https://www.youtube.com/watch?v=e0Dt5SRzLF8

Her Channel

kirbysuperstar posted:

I also have a friend who is a budding vtuber (she got twitch affilate back in November, though!) that I'll quickly plug if that's okay.

https://twitter.com/ZafraZaleska
https://www.twitch.tv/transienttiefling

Games, some drawing stuff, usually has a friend or two with her.

Sundae posted:

ShallNoiseUpon posted:

I hope Gura does a Calli rap.